Plaud

Tags

Install Options

Works With

Source

Skill Preview

# Plaud

Plaud ships two agent-facing surfaces from one vendor and one account:

- **`@plaud-ai/cli`** — terminal, scriptable, pipeable.
- **`@plaud-ai/mcp`** — Model Context Protocol server that auto-installs into Claude Code, Claude Desktop, Codex, Cursor/Windsurf/VS Code/Zed, plus an HTTP variant for Claude Web and ChatGPT.

Both authenticate via OAuth in the browser. Both live under `~/.plaud/` but use separate token files (`tokens.json` for CLI, `tokens-mcp.json` for MCP), so they can coexist.

## Skill structure

```
plaud/
├── SKILL.md              ← Router + install + quick reference (this file)
└── references/
    ├── cli.md            ← Full CLI command reference + config + env vars + exit codes
    ├── mcp.md            ← Full MCP install per client + manual JSON + HTTP variant + tool inventory
    ├── workflows.md      ← Cookbook: bulk export, date-range digest, cross-surface recipes
    └── troubleshooting.md ← Auth, token refresh, install, MCP "server disconnected"
```

## Choose a surface

| User intent | Use |
|-------------|-----|
| "Summarize Tuesday's standup" / "draft follow-up from this meeting" — conversational, inside an AI client | **MCP** |
| "Bulk-export every transcript from April to .txt files" — shell pipeline, cron, scripted | **CLI** |
| Working in Claude Web or ChatGPT (no local install possible) | **MCP via HTTP** (`https://mcp.plaud.ai/mcp`) |
| Headless / remote / CI machine | **CLI** with `PLAUD_*` env vars |
| Want both | Install both — they share auth but use separate token files |

When in doubt: if the agent should call Plaud on the user's behalf inside an AI client → MCP. If a shell script or human in a terminal needs the data → CLI.

## Install — MCP

Run once. The installer detects local AI clients, writes config, and opens the browser for OAuth:

```bash
npx -y @plaud-ai/mcp@latest install
```

Then **fully restart** the AI client (close and reopen the window — for Claude Code, exit and start a new `claude` session). Flags:

- `--yes` — auto-configure all detected clients, no prompts
- `--no-login` — skip browser sign-in (for remote/headless machines)

For clients not auto-detected, add this manually to their MCP config:

```json
{ "mcpServers": { "plaud": { "command": "npx", "args": ["-y", "@plaud-ai/mcp@latest"] } } }
```

For Claude Web / ChatGPT Web (HTTP variant): see `references/mcp.md`. Note: HTTP requests pass recording data through Plaud's US-hosted MCP server in transit.

The installer also loads Plaud's own built-in skills (`plaud-browse`, `plaud-find`, `plaud-read`, `plaud-digest`, `plaud-followup`, `plaud-export`) into Claude Code. **This skill defers to those for the in-MCP execution playbooks** — do not re-implement their logic here.

## Install — CLI

```bash
npm install -g @plaud-ai/cli
plaud login          # opens browser for OAuth
plaud me             # confirm sign-in
```

Requires Node.js ≥ 20.

## Quick reference

| Task | CLI command | MCP tool |
|------|-------------|----------|
| Sign in | `plaud login` | `login` |
| Whoami | `plaud me` | `get_current_user` |
| List recordings | `plaud files` / `plaud recent` / `plaud today` | `list_files` |
| Search by keyword | `plaud search "Q2"` | `list_files` with `query` |
| Filter by date | `plaud search ... --from 2026-04-01 --to 2026-04-30` | `list_files` with `date_from` / `date_to` |
| Get one recording | `plaud file <id>` | `get_file` |
| Audio download URL (24h) | `plaud audio <id>` | `get_file` → `presigned_url` |
| Transcript | `plaud transcript <id>` | `get_transcript` |
| AI summary | `plaud summary <id>` | `get_note` |
| Sign out | `plaud logout` | `logout` |

Pipe-friendly: CLI errors go to **stderr**, results go to **stdout**. `plaud transcript <id> -o file.txt` or `plaud summary <id> -o file.md` write to disk.

## Deep references

| Read when… | File |
|------------|------|
| Need a full flag table, exit codes, config file, or `PLAUD_*` env vars | `references/cli.md` |
| Setting up MCP in a non-default cl